A developer has released an open-source experimental framework called Gear to explore recursive self-improvement (RSI) and is seeking community feedback on its design.
The framework implements an iterative loop where a meta-agent reviews execution traces and evaluation results to propose changes to the task agent's instructions, tools, or workflows. The current implementation focuses on improving the agent's harness rather than direct model training, which remains experimental.
The author specifically requests advice on ensuring improvements generalize to new tasks without overfitting, potential simplifications to the loop design, and relevant learning resources.